LabsRun / inspect / learn

Change one variable. Keep the evidence connected.

The local runner handles trusted, offline, zero-cost capsules. The browser-first lesson replays the same seed canary without executing code. Neither surface is arbitrary public code execution.

Current surfaces02 available / distinct trust boundaries

A narrow runner.
A clear boundary.

Exists locallycnlab

Trusted-source local runner

It validates capsule paths, passes a minimal environment allowlist, uses shell=False, bounds subprocess and capsule wall time, captures outputs, snapshots sources, hashes artifacts, and performs two complete trials.

Offline, zero-cost, risk-tier-0 capsules only
Declared seeds and outputs verified
Source stability checked during execution
Bundle integrity recomputed

It is not a sandbox.The recorded network policy is not enforced by the operating system. Do not run untrusted submissions, learner code, secrets, personal data, or dangerous workloads.

Available locallybrowser replay

Seed-sensitivity learning lab

The browser-first exercise derives from the retained canary instead of inventing a second toy result. Select seeds, recompute descriptive statistics, and practice an evidence-safe claim.

01Select
02Compare
03Summarize
04Check a claim
05Read limits

Open the lesson

Replay only.The lesson uses embedded retained results. It executes no research code and earns no reproduction label. No public executor, notebook service, multi-user runtime, or server-side learner code is running.

GuardrailsNo arbitrary public execution

Learning without widening the trust boundary.

Use safe inputs

Synthetic or clearly licensed public data; never learner secrets.

Keep the capsule canonical

Claims and figures stay linked to the retained evidence version.

Make reset deterministic

Starter state, checks, solution, cleanup, and export are part of the lab.

Escalate real risk

Risk tier 2 requires isolated human-approved execution; autonomous tier 3 work is rejected.